The definition of machine translation (MT) has evolved over the years, possibly due to the transformation of computers into powerful data-processing machines (Tekwa Reference Tekwa2023). In this context, neural machine translation (NMT) emerged and gained widespread use, aided by the development of computer-assisted translation (CAT) tools and large language models (LLMs). In recent years, MT and NMT have been applied across various domains, such as education and software localisation. These technologies are examined from multiple perspectives, including translation quality, economic benefits, and the ethical and ecological consequences of MT and NMT. Some view MT and NMT as valuable tools due to their advancements and applications, while others remain sceptical about their potential social implications. Automating Translation is a seminal work in the series of Routledge Introductions to Translation and Interpreting, which explores both the strengths and concerns surrounding MT and NMT.

The book fills a significant gap in the field of MT by providing a comprehensive overview of its development stages and current research trends. It also offers valuable insights for future studies. In contrast to Towards Responsible Machine Translation: Ethical and Legal Considerations in Machine Translation (Moniz and Parra Escartín Reference Moniz and Parra Escartín2023), which emphasises the ethical dilemmas, legal risks, and social responsibilities arising from the practical application of machine translation, and Translation Tools and Technologies: A Practical Guide for Students and Translators (Rothwell et al. Reference Rothwell, Moorkens, Fernández-Parra, Drugan and Austermuehl2023), which focuses on practical operational guidelines for translation tools, Automating Translation successfully constructs a core knowledge system in the field of translation automation by thoroughly explaining the complete chain from data and models to applications and ethics. The book is structured into eleven chapters, each addressing key challenges and issues in MT. These chapters can be grouped into four sections: the first section (Chapters 1 and 2) introduces the fundamentals and data behind MT; the second section (Chapters 3–5) discusses the essential tools and technologies widely used in MT; the third section (Chapters 6–9) focuses on the practices and applications of NMT; and the final section (Chapters 10 and 11) explores the broader social and ethical implications of MT. The primary aim of the book is to present relevant topics on MT to readers, making it an excellent resource for students majoring in or interested in the development of translation technology, as well as for specialists in the field, including translators and professors.

In Chapter 1, titled The Roots of Machine Translation, Moorkens, Way and Lankford elaborate on the various stages of development in MT and the dynamic relationships between them. They provide a comprehensive analysis of the operational mechanisms and challenges of rule-based machine translation (RBMT). Additionally, they examine why statistical machine translation (SMT) surpassed RBMT to become the dominant paradigm. Although MT has achieved these advancements, the authors remind readers that translation remains a complex task, and MT continues to face numerous unresolved issues.

Chapter 2, Data for Machine Translation, focuses on translation data in MT, which consists of source texts and their corresponding translations. The authors point out that translation data in MT can be used to train data-driven translation models. They also introduce the concept of translation editing interfaces, which have gradually evolved into translation memory (TM) systems capable of storing large amounts of translation data. Translation data can also be obtained through open access resources, web scraping, and synthetic data generated by back-translation. However, translation data that is not produced by CAT tools often suffers from issues related to alignment and formatting. Furthermore, the ownership of such data can be difficult to ascertain. Chapter 2 also highlights the crucial role that the quantity of data plays in determining the quality of MT or NMT.

In Chapter 3, entitled Translation Memory and Computer-Assisted Translation Tools, the authors introduce three fundamental functions of CAT tools: TM, MT, and termbases. By the mid-2000s, many CAT tools had begun incorporating MT services, often via application programming interfaces (APIs), while more recent developments have seen the integration of generative AI technologies. TM, which serves as the core of CAT tools, is a repository of translation pairs, where the source text and the target text are matched in units known as translation units (TUs). CAT tools can automatically identify and create TUs, allowing for the efficient retrieval of similar or identical source segments from the TM, thereby improving translation efficiency. Unlike TM, MT generates translations rather than merely retrieving pre-existing translation segments. Additionally, termbases play a crucial role in translation projects involving specialised terminology. It is also important to note that the advent of translation platforms has shifted the ownership of translation resources from individual translators to platform owners.

In Chapter 4, the authors provide a comprehensive understanding of neural networks (NN) and NMT. NN evolved from feed-forward neural networks (FFNN) to recurrent neural networks (RNN) and ultimately to the transformer architecture. After explaining the role of back-propagation in NN training, the authors distinguish FFNN from RNN by emphasising the recurrent connections of the latter. As a critical component in natural language processing (NLP) systems, RNNs allow nodes to connect with themselves, meaning that previous outputs can be used as inputs. Regarding the transformer architecture, the authors hold that the attention mechanism enables the model to focus on different parts of the input, thus improving translation accuracy and coherence. The evolution in NN technologies paved the way for the development of NMT, which has evolved from training MT models to delivering translation services based on probabilistic mathematics. In NMT, input and output sentences are treated as sequences, leading to the creation of sequence-to-sequence models. Despite its advances, NMT still faces several challenges, including issues with terminology, domain-specific outputs, and hallucinations.

In Chapter 5, titled Machine Translation Evaluation, the authors discuss various methods of machine translation evaluation (MTE). Since the 1950s, MTE has played a significant role in assessing the quality of machine translation, though it remains both complex and controversial. MTE is typically divided into two main types: automatic evaluation and human evaluation. Automatic evaluation metrics (AEMs) can be categorised into two types: “precision-based metrics,” such as Bilingual Evaluation Understudy (BLEU), METEOR, and NIST, and “error-based metrics,” such as translation edit rate (TER) and word error rate (WER). Most AEMs assess the quality of machine translation based on its similarity to human translation. Human evaluation, on the other hand, includes methods like error annotation and post-editing, inter-annotator agreement, crowdsourcing evaluation, and quality estimation. Based on the discussion of MTE, the authors conclude that the optimal approach to translation evaluation combines both human and automatic evaluation.

Chapter 6, entitled Neural Machine Translation: Build or Buy?, continues the discussion of NMT introduced in Chapter 4. Moorkens, Way, and Lankford present several key NMT toolkits, including OpenNMT, Marian-NMT, Joey NMT, ModernMT, OPUS-MT, MutNMT, and adaptNMT. Notably, unlike other NMT toolkits, adaptNMT does not rely on a command-line interface, making it particularly suitable for educational and research environments due to its ease of use and lower carbon footprint. Additionally, they proceed to explain the components of adaptNMT in detail, covering its architecture, user interface, initialisation and logging, modes of operation (including local and cloud-based options), model customisation, subword segmentation, and translation evaluation. Moorkens, Way, and Lankford also provide a case study demonstrating how adaptNMT can be used for training models from datasets, focusing on infrastructure, approach, and architecture adjustments. This case highlights the potential of machine translation to bridge linguistic gaps in key domains.

In Chapter 7, titled Building Machine Translation Models with Colab, the authors focus on Google Colab, which provides access to free computational resources such as GPUs and Tensor Processing Units (TPUs). In comparison to Colab, Colab Pro and Pro+ offer enhanced accessibility and additional resources, though they are not free. Currently, Google Colab supports AI coding features, making it possible to train machine translation models by using the OpenNMT toolkit within a basic Colab notebook. Colab notebooks can also be shared, including text, code, output, and comments. The general workflow includes connecting Google Drive, installing the MT engine, verifying GPU availability, building a vocabulary, and starting the training process. Additionally, Gradio can be used to create user interfaces for MT models in Colab.

In Chapter 8, titled Machine Translation Post-Editing, the authors explore the post-editing aspect of MTE discussed in Chapter 5. Post-editing has evolved alongside the history of MT, from RBMT to SMT and NMT. For post-editing, translators need to follow some specific guidelines. Post-editing is typically carried out in CAT tools. Hybrid post-editing refers to MT and its post-editing occurring simultaneously. Over time, custom post-editing tools and Interactive Translation Prediction (ITP) have been developed and integrated into commercial CAT tools. Despite the rapid advancements in MT quality, some translators remain sceptical about post-editing due to concerns about the effort required.

Chapter 9 discusses the applications of MT in software localisation, game localisation, website localisation, and audiovisual translation (AVT) for subtitling and dubbing. In the field of software localisation, NMT is widely used due to its lower cost and higher speed. However, NMT often requires the support of CAT tools to align with the specific terminology of localisation projects. In-game localisation and custom MT or NMT, along with CAT tools, are used to help ensure playability and immersion. Compared to software and game localisation, web localisation tends to be easier because of the separation of content and structure. However, web localisation has a higher degree of internationalisation, meaning that content, text, and images are often culture-specific. For AVT, MT has become an indispensable tool, especially for subtitling. Additionally, AVT provides a pathway for draft dubbing scripts, which are more complex than subtitling. Notably, some educational companies have successfully completed fully automated dubbing using MT.

In Chapter 10, Large Language Models and Multilingual Language Models: The Future of Machine Translation?, the authors focus on LLMs and Multilingual Language Models (MLLMs), as well as methods for fine-tuning these models. Common LLMs include GPT-J, GPT-4, Microsoft Copilot, Gemini, BERT, DeepSeek, and Hugging Face. Interaction with LLMs can be easily achieved using simple prompts. One of the most innovative AI projects in MLLMs is No Language Left Behind (NLLB) (Costa-jussà et al. Reference Costa-jussà, Cross, Çelebi, Elbayad, Heafield, Heffernan, Kalbassi, Lam, Licht, Maillard and Sun2022), which aims to address the challenge of language inclusivity in AI. The starting point of NLLB is NLLB-Seed, a model that allows for minimal adjustments for a specific language. NLLB also has variants, such as NLLB-200-600 M and NLLB-200-1.3B. When constructing (M)LLMs, key considerations include designing and training the model architecture. Tools like ChatGPT can be used to develop translation tools. Additionally, they also highlight the potential of LLMs, including applications in language translation (Costa-jussà et al. Reference Costa-jussà, Cross, Çelebi, Elbayad, Heafield, Heffernan, Kalbassi, Lam, Licht, Maillard and Sun2022), chatbots, virtual assistants, and other creative uses.

In Chapter 11, entitled Sociotechnical Effects of Machine Translation, the authors delve into the ethical issues surrounding NMT and LLMs, highlighting the potential risks of MT for both environmental and social sustainability. To address this theme, some developers of MT systems have worked on increasing efficiency to reduce electricity consumption. While environmental sustainability issues related to MT are important, the relationship between MT and social sustainability has broader implications. This includes concerns about translator unemployment, potential risks for MT users, and ongoing debates about copyright and translation data. Nevertheless, MT still plays a positive role in serving the public interest. For instance, MT can offer automatic translation services in times of crisis, such as the rapid dissemination of information among different countries during the COVID-19 pandemic.

Overall, the emergence of AI, which relies on LLMs, has extended its reach from translation to various other domains within intellectual labour. MT is one of the most important applications of AI in the field of translation. This book, Automating Translation, introduces new topics and perspectives to examine the development and effects of MT and NMT and also covers multiple aspects concerning the application and evaluation of MT and NMT. However, some important but unexplored aspects remain, for example, the topic of “cultural literacy”. The term “cultural literacy” refers to the basic information needed to thrive in the modern world (Hirsch Reference Hirsch1988). In the context of MT, “cultural literacy” pertains to how MT systems consider intercultural factors during translation and the ability of MT to conduct literary research. With the rapid development of MT, more researchers and translators have begun to explore its potential for literary translation and related research. Therefore, integrating “cultural literacy” into MT remains one of the biggest unsolved challenges. This book has introduced different translation tools or CAT tools, such as DeepL to assist literary translators. However, these tools can only achieve “surface translation” of literary contexts, sometimes even causing misunderstandings about cultural aspects. Despite these limitations, the book provides valuable insights for undergraduate and research students interested in MT. It highlights the challenges and issues of MT and NMT and suggests possible solutions.